Sandra Loosemore wrote:
maybe this one is simple
enough that I can make more progress on it. It's probably something
being added to the wrong gimple_seq somewhere.
Good luck!
* * *
I played around without fully checking the the validity. I found
an unrelated ICE - but also was puzzled by "‘i’ is used uninitialized"
warnings.
* * *
iter-test2.c:14:11: internal compiler error: base pointer cycle detected
14 | #pragma omp target map(to: x[:4]) map(iterator(i=low:high:stride),
to: x[i].ptr)
| ^~~
0x27fb3af internal_error(char const*, ...)
gcc/diagnostic-global-context.cc:787
0xec1444 omp_tsort_mapping_groups_1
gcc/gimplify.cc:11438
--------<cut>-------------
struct array_ptr
{
int *ptr;
};
void
f (struct array_ptr *x, int low, int high, int stride)
{
#pragma omp target map(to: x[:4]) map(iterator(i=low:high:stride), to:
x[i].ptr)
{
__builtin_printf ("f: dev %p / %p\n", x[0].ptr, x[2].ptr);
}
}
--------<cut>-------------
It works if I add 'x' [namely: 'map(to: x, x[:4])'] - but doing so gives with
-Wall:
iter-test2.c:14:11: warning: ‘<anonymous>’ is used uninitialized
[-Wuninitialized]
14 | #pragma omp target map(to: x, x[:4]) map(iterator(i=low:high:stride),
to: x[i].ptr)
| ^~~
iter-test2.c:14:78: note: ‘<anonymous>’ was declared here
14 | #pragma omp target map(to: x, x[:4]) map(iterator(i=low:high:stride),
to: x[i].ptr)
|
^
Or when doing:
void
f (struct array_ptr *x, int low, int high, int stride)
{
#pragma omp target map(to: x) map(iterator(i=low:high:stride), to:
x[i].ptr[:2])
iter-test.c: In function ‘f’:
iter-test.c:15:71: warning: ‘i’ is used uninitialized [-Wuninitialized]
15 | #pragma omp target map(to: x) map(iterator(i=low:high:stride), to:
x[i].ptr[:2])
| ^
iter-test.c:15:46: note: ‘i’ was declared here
15 | #pragma omp target map(to: x) map(iterator(i=low:high:stride), to:
x[i].ptr[:2])
* * *
However, the original issue is independent as the following is unrelated to
iterator and also gives the ICE:
#pragma omp target map(to: x[:4]) map(to: x[0].ptr[:2], x[2].ptr[:2])
Still, the 'i' is uninitialized is unexpected :-/
* * *
I had expected that the following works, but it fails at runtime with:
void
g (struct array_ptr *x)
{
__builtin_printf ("g: host %p / %p\n", x[0].ptr, x[2].ptr);
#pragma omp target map(to: x, x[:4], x[0].ptr[:2], x[2].ptr[:2])
__builtin_printf ("g: dev %p / %p\n", x[0].ptr, x[2].ptr);
}
but gives
g: host 0x7ffc79878128 / 0x7ffc79878118
libgomp: Trying to map into device [0x7ffc798780f0..0x7ffc79878110) object
when [0x7ffc79878100..0x7ffc79878108) is already mapped
for
main ()
{
int x1[] = {1, 2};
int x2[] = {11, 22};
int x3[] = {111, 222};
int x4[] = {1111, 2222};
struct array_ptr var[4];
var[0].ptr = x1;
var[1].ptr = x2;
var[2].ptr = x3;
var[3].ptr = x4;
g(var);
}
Tobias
PS: Just to dump the findings; I am too tired to analyse this this evening
and to decide what's valid OpenMP and what's not.
